We explore the way that the flexibility inherent in the lexicon might be incorporated into the process by which an environmentally grounded artificial agent-a robot-acquires language. We take flexibility to indicate not only many-to-many mappings between words and extensions, but also the way that word meaning is specified in the context of a particular situation in the world. Our hypothesis is that embodiment and embededness are necessary conditions for the development of semantic representations that exhibit this flexibility. We examine this hypothesis by first very briefly reviewing work to date in the domain of grounded language learning, and then proposing two research objectives: 1) the incorporation of high-dimensional semantic representations that permit context-specific projections, and 2) an exploration of ways in which non-humanoid robots might exhibit language-learning capacities. We suggest that the experimental programme implicated by this theoretical investigation could be situated broadly within the enactivist paradigm, which approaches cognition from the perspective of agents emerging in the course of dynamic entanglements within an environment.